Creative Minds Development is a creative tech studio building AI, cloud and cybersecurity systems for African enterprises — we bring ideas to life.
Joined May 2026
- Tweets 253
- Following 44
- Followers 6
- Likes 20
40 Photos and videos
Pinned Tweet
Jun 5
We cancelled our Upwork account today.
A decade on the platform since 2016. Thousands of dollars in Connects. Hundreds of proposals.
Top-slot auction bids. Zero contracts.
Not one.
Here is the math no one explains before you start.
1
1
1
25
Most teams reach for fine-tuning two phases too early.
Before customisation, try in this order: 1. Better prompts (days of work) 2. Better RAG (chunking, embedding, vector store, hybrid weights, rerank) 3. Better tool design (narrower tools, sharper OpenAPI descriptions)
THEN consider customisation.
2
1
5
Distillation is the most ops-impactful path for 2026:
Claude as teacher → generate training data → distil into Llama 4 Scout or Titan Lite → near- Claude quality at student cost.
Pays for itself in months on high-volume work.
1
7
Eval harness is non-negotiable.
Every customisation decision is faith-based without it. The held-out test set, the LLM-as-judge calibration, the rollback gate.
Full piece: creativeminds.dev/blog/model…
#BedrockAgents #FineTuning #LLMOps #ClaudeAI #ModelDistillation
2
We posted the Upwork critique a week ago.
Across X, LinkedIn, Reddit, and HN, the replies broke into five patterns. Worth sharing because they explain why the platform has so few public critics despite the math being obvious.
1
7
Pattern 3: “Maybe just don’t bid on bad jobs.”
The platform’s product design makes the bad jobs indistinguishable from the good ones at application time. Verification only happens AFTER you spend the Connects. The asymmetry is the point.
1
3
Pattern 4: “It’s not perfect but what’s the alternative?”
The alternatives are: direct outreach, published work, specialist networks, your own brand. None of them require buying Connects to apply.
Pattern 5: silence from people still on the platform who don’t want to be seen criticising it. We were one of those. Not anymore.
creativeminds.dev/blog/upwor…
#Upwork #Freelancing #GigEconomy #FreelanceCommunity #PayToPlay
2
Jun 13
Bedrock Agents (managed loop) is the right answer for ~60-70% of production agents: singletask, well defined, AWS-native.
The other 30-40% need control over the loop. Three credible paths.
1
1
7
Jun 13
What earns Strands the centre: steering handlers.
A blocking guardrail says “no.” A steering handler says “no, do it this way instead.” Agent reads it, corrects, proceeds.
Their benchmark: prompt-only agents recovered from 82.5% of errors. Steering handler agents recovered from every one.
1
5
Jun 13
Plus hooks for audit, Agent-as-Tool, Swarm, conversation managers, event.interrupt() for HIL on destructive actions.
Full piece: creativeminds.dev/blog/open-…
#StrandsAgents #LangChain #BedrockAgents #AgentCore #OpenSource
2
Jun 12
Most production RAG failures aren’t model failures.
They are failures of one of five gates upstream of the model: source storage → chunking → embedding → vector storage → retrieval.
The model dutifully synthesizes bad context. The Claude session looks competent. The answer is wrong.
1
1
15
Jun 12
Vector store choice has the longest reversal cost. Switching means re-ingesting everything.
OpenSearch Serverless as AWS-native default. pgvector for moderate scale on existing Postgres. Pinecone for existing investment.
Choose deliberately.
1
32
Jun 12
Hybrid search (BM25 vector) Cohere Rerank metadata filters from the user’s claims.
The metadata filter is also the access-control boundary — see RBAC for KBs piece.
Full architecture: creativeminds.dev/blog/rag-w…
#RAG #BedrockAgents #VectorDB #ClaudeAI #GenAI
4
Jun 11
Upwork in 2016: the platform takes a commission when a freelancer completes work.
Freelancer and platform interests align — both want the deal to close.
Upwork in 2026: the platform monetises the attempt to find work. Application fees, boost auctions, ghost-listing revenue.
We were there before the inversion.
1
17
Jun 11
This is not unique to Upwork. It is the pattern.
Marketplaces start by aligning with the small side of the network (the supply — freelancers, sellers, drivers).
They end by extracting from the small side, because the small side is captive.
Upwork is the textbook case.
1
5
Jun 11
A decade in. Thousands of dollars in. Zero contracts.
We cancelled the account.
If you’ve been on Upwork since the commission days and watched the same trajectory, you already know what this article is going to say. It’s at creativeminds.dev/blog/upwor….
#Upwork #PlatformEconomics #GigEconomy #Freelancing #MarketplaceCritique
2